Deep deterministic policy gradient and graph convolutional networks for topology optimization of braced steel frames

Abstract

We propose a method for topology optimization of braced frames under static seismic loads using Deep Deterministic Policy Gradient (DDPG) and Graph Convolutional Network (GCN). The structure is interpreted as a graph where structural elements and element configurations are represented by the node feature matrix and adjacency matrices, respectively. Using this graph representation, the DDPG agent with GCN architecture can observe the properties of the frame, and make the decision to either add braces into the frame or enlarge sections of frame elements by selecting from a list of available sections. During the optimization process, the initial structure that cannot withstand the seismic load is modified by the agent until all constraints are satisfied. The trained agent can be applied to frames of different sizes and can obtain competitive results with less computational cost compared to the genetic algorithm.